Vibe Coding: The Future of Software Development or a Passing Fad? Insights from Y Combinator

hat if coding wasn’t about sweating over syntax or debugging line-by-line, but about embracing "the vibes" and letting AI do the heavy lifting? That’s the essence of Vibe Coding—a term coined by Andrej Karpathy in a viral post—and it’s taking the startup world by storm.

Recently, the partners at Y Combinator (YC)—Gary, Jared, Harj, and Diana—sat down for an episode of The Lightcone to unpack this phenomenon, drawing insights from the current YC batch of founders. What they found is nothing short of a revolution in how software is built.

What is Vibe Coding?

Karpathy describes Vibe Coding as a mindset: "fully give in to the vibes, embrace exponentials, and forget that the code even exists." It’s coding at warp speed, powered by AI tools like Cursor, Claude, and o3, where the focus shifts from writing every line to orchestrating outcomes.

The YC partners surveyed their current batch—founders of highly technical startups—and discovered that this isn’t just a quirky trend. It’s becoming the dominant way to code.

One founder summed it up: “I’m no longer an engineer—I’m a product person.” Another said, “I don’t write code much; I just think and review.”

These aren’t non-technical dreamers; they’re builders who, a year ago, would’ve crafted every line themselves. Now, AI generates upwards of 95% of some codebases. Yes, you read that right—95%.

The Models People Are Using Now

The tools behind Vibe Coding are evolving at breakneck speed, and the YC batch reflects this dynamic landscape.

Cursor reigns supreme, a favorite since summer 2024 when its adoption surged among founders. “It’s by far the leader,” Diana noted, citing its ability to turbocharge coding speed.

But there’s a new contender: Windsurf. Jared highlighted its edge: “Cursor needs you to tell it which files to look at in your codebase. Windsurf indexes the whole thing and figures it out on its own.” For large projects, that’s a game-changer.

Not far behind, ChatGPT remains a staple—not for raw code generation, but for tapping into reasoning models like o1 and o3. Founders lean on it for debugging, where precision trumps speed.

“Cursor and Windsurf are still in the pre-reasoning world,” Diana explained, referencing a time before test-time compute models (think less than six months ago). For thornier problems, ChatGPT’s firepower fills the gap.

Other players? Devon gets a nod but falters on serious features, lacking deep codebase awareness—it’s relegated to small tasks.

Meanwhile, Claude Sonnet 3.5 holds strong as the “big game in town” from six months back, though it’s now neck-and-neck with reasoning upstarts like o1, o1 Pro, and o3. DeepSeek R1 is emerging as a viable contender, and Gemini sneaks in for its massive context window.

“Some founders dump their entire codebase in Gemini and say, ‘Fix this,’” Jared shared. It’s hit-or-miss, but when it works, it’s a one-shot wonder.

Self-hosting is also on the rise, especially for founders guarding sensitive IP. And keep an eye on Flashback 2.0—its long context plus reasoning could shake things up, though adoption is still early.

What Percentage of Code is Being Written by LLMs?

Here’s where it gets wild. The YC team asked founders: “What percent of your codebase is AI-generated?” Not libraries or imports—just the characters typed by humans versus emitted by LLMs.

The answer? One-quarter of founders said over 95% of their code is AI-written. “That’s an insane statistic,” Jared marveled.

These aren’t non-technical dabblers. “Every one of these people is highly tactical, completely capable of building their product from scratch a year ago,” he added.

Yet today, AI dominates their workflows. Some founders—especially younger ones who learned to code in the last two years—don’t even know a pre-Cursor world.

“One of my best companies this batch is exactly this,” Gary said. “Extremely technical minds, not classically trained in CS, but incredibly productive. AI writes almost the entire thing.”

Think of them as AI natives, akin to digital natives who grew up with the internet. With degrees in math or physics, not programming, they wield system-thinking skills and vibe their way to stellar products.

“Coding bootcamps used to struggle to retrain physicists into programmers,” Jared recalled. “It took too long to learn syntax and libraries. Now, it’s a new world.”

What’s Changed, What’s Stayed the Same?

This shift turbocharges productivity for non-traditional coders, but it’s not all vibes and exponentials. AI excels at speed—95% of a codebase in a founder’s hands overnight—but struggles with debugging and systems-level thinking.

“Vibe Coding enables tactical minds from other disciplines to become programmers fast,” Gary noted.

Yet, as companies scale, the need for architects who grasp the bigger picture persists.

Hiring hasn’t caught up either. Whiteboard algo-tests feel quaint when LLMs can ace them. Companies like Stripe and Gusto once pioneered “build a to-do app in three hours” interviews, prioritizing output over theory.

Now, even that may be outdated. “Maybe raw coding speed isn’t the metric anymore,” Jared mused. Taste and systems savvy might be the new gold.

Zero-to-One vs. One-to-N

This duality mirrors startup growth. Vibe Coding shines in the zero-to-one phase—shipping features fast to find product-market fit. Founders like Yoav from Cix juggle two Cursor windows, prompting parallel features. But scaling to millions of users?

That’s where classical training kicks in. Think Facebook’s PHP-to-HipHop pivot or Twitter’s battle with Ruby and spiky Super Bowl traffic. “You’re so lucky to get to one,” Harj said, but beyond that, you need architects who won’t flinch at bare metal.

The AI-Native Generation

A fascinating twist: some founders in this batch never knew a world without AI coding tools. Degrees in math or physics, not CS, they’re “AI natives” who skipped classical training and still ship world-class products.

“The barrier to entry is so low,” Diana observed. Bootcamps once struggled to retrain physicists into coders—now, AI levels the field overnight.

But taste and systems thinking still matter. “You need deliberate practice to be exceptional,” Jared argued, citing Picasso’s journey from lifelike sketches to abstract genius.

Good-enough engineers will flood the market, but the top 1%—the outliers—will blend vibe-coding speed with deep understanding.

Hiring in the Vibe Era

How do you hire for this? Traditional whiteboard interviews feel archaic when AI can ace tic-tac-toe in seconds. Jared, who co-founded Triplebyte and conducted thousands of technical interviews, suggested a rethink: test for tool fluency, debugging grit, and product taste.

“Can they call out bad code—human or AI?” he asked. Gary added a founder’s lens: “If you’re not technical enough to spot bullshit, you’re sunk.”

Not a Fad—It’s Here to Stay

The YC partners agree: Vibe Coding isn’t going away. “If you’re not doing it, you might just be left behind,” they warned.

A quarter of their batch says AI writes 95%+ of their code, and that’s not outliers—it’s the new norm. Debugging and systems challenges persist, but the trajectory is clear: exponential acceleration is reshaping software engineering.

So, are you ready to embrace the vibes? Because the beanstalk’s already growing—and it’s not stopping anytime soon.